LLM+ASP: Self-Correction Enables Task-Agnostic Nonmonotonic Reasoning

A novel framework named LLM+ASP has been developed to convert natural language into Answer Set Programming (ASP), which is a nonmonotonic formalism grounded in stable model semantics. This innovation allows large language models (LLMs) to engage in defeasible reasoning without the need for task-specific engineering. In contrast to earlier methods that depend on monotonic logics like SMT, which fall short in representing defeasible reasoning, LLM+ASP provides a consistent approach across various reasoning tasks. The system incorporates an automated self-correction feature to enhance logical coherence and lower computational expenses. The study, available on arXiv (2604.27960), points out that existing neuro-symbolic techniques often necessitate manually created knowledge modules or specific prompts, which LLM+ASP circumvents.